atel 0.1.0

Browser-based GLM workbench for actuarial pricing

Atelier

Browser-based GLM workbench for actuarial pricing

Build, fit, diagnose, and iterate on Generalized Linear Models — without leaving your browser.

---

Why Atelier?

Actuarial pricing teams deserve modern tooling. Atelier wraps rustystats — a high-performance Rust-backed GLM engine — in a clean, interactive UI designed for the model-building workflow.

• One command — uv add atel, run atel, and you're in the browser. No Docker, no Node.js, no config files.
• Fast — Rust-powered fitting is 5–20x faster than statsmodels. Iterate in seconds, not minutes.
• Visual — Factor charts, actual-vs-expected, lift curves, diagnostics, and model comparison — all live-updating as you build.
• Transparent — Every model action generates runnable Python code. Copy it, version it, schedule it.
• Local — Your data never leaves your machine. No accounts, no cloud, no telemetry.

---

Installation

uv add atelier
# or
pip install atel

Requires Python 3.13+. Installs everything — backend, frontend, engine. No separate build steps.

Quick start

atel                  # starts server, opens browser
atel --port 9000      # custom port
atel --no-browser     # headless / server-only

The atelier command works too - atel is just shorter.

---

Features

Model building

• 8 GLM families — Gaussian, Poisson, Binomial, Gamma, Tweedie, Quasi-Poisson, Quasi-Binomial, Negative Binomial
• Rich term types — categorical, linear, B-splines, natural splines, target encoding, frequency encoding, expressions
• Monotonic constraints — enforce increasing/decreasing effects on splines and linear terms
• Interactions — standard product terms, target-encoded interactions, frequency-encoded interactions
• Regularization — Ridge, Lasso, Elastic Net with cross-validated alpha selection
• Train/test split — holdout validation with stratified splitting

Diagnostics

• Factor-level A/E — actual vs expected charts for every factor, fitted or not
• Score tests — chi-squared significance for candidate factors before fitting
• Lift charts — Gini, AUC, KS statistics with decile breakdown
• Calibration — Hosmer-Lemeshow test, decile calibration with confidence intervals
• Residual analysis — deviance, Pearson, and working residuals
• VIF & multicollinearity — variance inflation factors with severity coloring
• Model comparison — side-by-side metrics against a base model

Data exploration

• Pre-fit analysis — response distribution, zero inflation, overdispersion detection
• Correlation matrix — numeric correlations and Cramér's V for categoricals
• Shape detection — automatic recommendations for term types per factor
• Interaction detection — greedy residual-based search for potential interactions

Workflow

• Model history — every fit is saved and versioned, compare any two iterations
• Context menus — right-click factors to change term types, toggle inclusion, reorder
• Keyboard-driven — search, filter, and navigate without reaching for the mouse
• Code generation — live Python script panel mirrors your UI actions

---

How it works

┌─────────────────────────────────────┐
│  Browser (React + TypeScript)       │
│  ┌──────────┬──────────┬──────────┐ │
│  │  Spec    │ Results  │  Code    │ │
│  │  Panel   │  Panel   │  Panel   │ │
│  └──────────┴──────────┴──────────┘ │
│           REST + WebSocket          │
└──────────────┬──────────────────────┘
               │
┌──────────────┴──────────────────────┐
│  Python backend (FastAPI + uvicorn) │
│  ┌────────────────────────────────┐ │
│  │  rustystats (Rust GLM engine)  │ │
│  └────────────────────────────────┘ │
│  Storage: ~/.atelier/ (SQLite)      │
└─────────────────────────────────────┘

Atelier is a single Python package. The React frontend is pre-built and bundled as static assets inside the wheel — no Node.js runtime required. The backend serves the UI and exposes a REST + WebSocket API for model fitting and real-time progress updates.

---

Development

git clone https://github.com/PricingFrontier/atelier.git
cd atelier
uv sync

# Backend (API server)
uv run atel --no-browser --port 8457

# Frontend (separate terminal, with hot reload)
cd frontend && npm install && npm run dev

Open http://localhost:5173 for the Vite dev server with HMR proxying API requests to the backend.

Running tests

uv run pytest

Building the frontend for production

cd frontend && npm run build
# Output lands in src/atelier/static/

---

Tech stack

Layer	Technology
GLM engine	rustystats (Rust)
Backend	FastAPI, uvicorn, SQLAlchemy, polars
Frontend	React 19, TypeScript, Vite, Tailwind CSS, shadcn/ui
Charts	Plotly.js, Recharts
State	Zustand
Storage	SQLite (via aiosqlite) at ~/.atelier/

---

License

Eclipse Public License 2.0